Only released in EOL distros:
Package Summary
This package provides a detector for cabinet fronts from dense depth images (as produced by Microsoft's Kinect sensor or the PR2's projected texture stereo system). The basic approach is to segment each image into planes and to search for the rectangle with the maximum support.
- Author:
- License: BSD
- Repository: alufr-ros-pkg
- Source: svn https://alufr-ros-pkg.googlecode.com/svn/tags/stacks/articulation/articulation-0.1.3
Package Summary
This package provides a detector for cabinet fronts from dense depth images (as produced by Microsoft's Kinect sensor or the PR2's projected texture stereo system). The basic approach is to segment each image into planes and to search for the rectangle with the maximum support.
- Author:
- License: BSD
- Source: svn http://alufr-ros-pkg.googlecode.com/svn/trunk/articulation
Contents
Please see the tutorials for a demo of this package using Microsoft's Kinect sensor.
Introduction
This package provides a module that detects rectangles in depth images and estimates their pose. Further, the module can track these rectangles over time and publish them as ArticulatedTrack messages. From these tracks, the articulation_models/model_learner_msg can learn the kinematic model underlying the observed motion.
A demo launch file is available that runs the detector, model learner, and visualization using RVIZ (see package tutorials). This implementation is very efficient: it runs on a Intel Core2 Duo with more than 10fps.
Approach
The detector itself works as follows:
- Input: a stream of dense depth images
- Segment the depth image using RANSAC
- Initialize a rectangle candidate in each plane
- Iteratively optimize the pose of the candidate by maximizing an objective function (see paper) until convergence
- Accept or reject the sample based on the pixel precision and recall
- Assign the detected rectangle to an existing track or initialize a new track
Output: a stream of ArticulatedTrack messages
Screen shots
The following images demonstrate how the detector, tracker, and model learning works.
Motivating example
Here, the lid of a laptop is opened and closed. By tracking the pose of the laptop lid, the kinematic model of the lid is estimated from its motion and visualized using RVIZ.
|
|
|
The lid of a laptop is opened and closed |
Visualization of the plane segmentation and the detected rectangle |
Resulting kinematic model learned from a sequence of observations |
Robustness against occlusions
Our approach also allows that parts of the tracked object are occluded:
|
|
Iterative pose fitting
Our approach segments the depth image into planes, and iteratively fits a rectangle to each plane. During pose fitting, our approach minimizes a cost function that is based on the number of true positive, false positive, and true negative pixels.
|
|
|
Segmentation and Detection
After fitting rectangles to all detected planes, our approach filters the candidates based on the pixel precision and recall. In these images, the detected planes are colored in blue/green/red and the detected rectangular objects are indicated using thick white borders.
|
|
Examples of learned kinematic models
|
|
|
|
More Information
More information (including videos, papers, presentations) can be found on the homepage of Juergen Sturm.
Report a Bug
If you run into any problems, please feel free to contact Juergen Sturm <juergen.sturm@in.tum.de>.
References
Jürgen Sturm, Kurt Konolige, Cyrill Stachniss, Wolfram Burgard. Vision-based Detection for Learning Articulation Models of Cabinet Doors and Drawers in Household Environments. In Proc. of the IEEE International Conference on Robotics and Automation (ICRA), Anchorage, USA, 2010. pdf
